Modeling
Summary + sample · full document is 1,931 words
player-expected-value.md covers the action-valuation half of player quality (xG, xA, xT, VAEP, EPV, OBV, off-ball valuation, defensive metrics). This note covers the rest of section A: latent-skill rating models, position-specific quality, aging, transfer-market priors, cross-league strength, and role/style decomposition.
The two notes together form Area A.
xV-style metrics value actions. A latent-skill rating values the player — a single (or low-dimensional) quantity meant to be portable across opponents, contexts, and seasons. Football's player-rating literature borrows heavily from basketball and chess.
Treat each player as a latent skill θ_p with a prior, observe minutes-weighted action values, fit by MCMC or variational inference. Partial pooling shrinks low-minutes players to position-and-age priors. The natural framing for a "career-stable" rating system. Few public implementations for football specifically; the framework is borrowed from baseball (Albert/Bennett-style models) and chess (Glickman).
Originally an NBA tool. Regress goal difference per minute on a sparse design matrix where each player gets a +1 (on the pitch) or 0 column. Coefficients are per-player rating impacts. Two issues, both severe in football:
…
Pro
Player Quality — Landscape Note (Area A, companion) runs 1,931 words. Pro members get every research note in full, plus the arbitrage feed, model outputs, and weekly updates.